Multi-Label Printed Chinese Character Attribute Recognition via Max-Pooling Enhanced Lightweight Networks

With the advancement of digital office automation, the automatic recognition of printed Chinese character attributes has significant applications in document processing and electronic archiving. However, existing studies primarily focus on font classification, while the rapid joint recognition of multiple attributes, such as bold, italic, and underline, remains a challenge. To address this issue, this paper constructs a printed Chinese character attribute dataset (PCCA) using editable documents and proposes an end-to-end multi-label recognition model (MP-HanNet) for printed Chinese character attributes. Based on an analysis of the characteristics of Chinese character images, we introduce max pooling as the core feature extraction operation and optimize the feature learning process by incorporating max pooling in the downsampling layers and parallel branches at different stages of the network. Additionally, max pooling and global average pooling (GAP) are integrated to construct a multi-label classification module. Experimental results on the PCCA dataset demonstrate that MP-HanNet achieves a better balance between accuracy and computational efficiency compared to mainstream lightweight networks. Furthermore, it exhibits stronger generalization capabilities on test images acquired through different methods (scanning and photography). This research provides new technical support for intelligent document processing and can be further extended to other multi-label image recognition tasks.